메뉴
Bootstrap AI coding stack: deploy LiteLLM proxy (via LiteLLM-Huawei-MaaS-Proxy skill), install opencode + oh-my-opencode-slim, mint virtual key, wire everything. Supports multi-key MaaS load balancing. TRIGGER on: opencode + Huawei MaaS setup, full-stack bootstrap, oh-my-opencode-slim-huawei-maas, deploy-litellm.
| name | oh-my-opencode-slim-huawei-maas |
| description | Bootstrap AI coding stack: deploy LiteLLM proxy (via LiteLLM-Huawei-MaaS-Proxy skill), install opencode + oh-my-opencode-slim, mint virtual key, wire everything. Supports multi-key MaaS load balancing. TRIGGER on: opencode + Huawei MaaS setup, full-stack bootstrap, oh-my-opencode-slim-huawei-maas, deploy-litellm. |
Bootstrap a complete AI coding stack on a single host: deploy LiteLLM proxy (via LiteLLM Huawei MaaS Proxy), install opencode with oh-my-opencode-slim plugin, mint a scoped virtual key, configure end-to-end. Supports multi-key MaaS load balancing for increased throughput. Idempotent — safe to re-run.
Both skills must be installed at these exact paths:
/home/
├── LiteLLM-Huawei-MaaS-Proxy/ ← LiteLLM proxy deployment
└── oh-my-opencode-slim-huawei-maas/ ← this project
These paths are enforced — no search, no fallback. The LiteLLM proxy is always at /home/LiteLLM-Huawei-MaaS-Proxy.
Both skills are extracted from the monorepo https://github.com/binrogithub/1-3-Cloud-Adoption-Skills.git:
MONOREPO="https://github.com/binrogithub/1-3-Cloud-Adoption-Skills.git"
TEMP_DIR="/home/1-3-Cloud-Adoption-Skills"
git clone --depth 1 "$MONOREPO" "$TEMP_DIR"
cp -r "$TEMP_DIR/AI/AI-Coding/LiteLLM-Huawei-MaaS-Proxy" /home/LiteLLM-Huawei-MaaS-Proxy
cp -r "$TEMP_DIR/AI/AI-Coding/oh-my-opencode-slim-huawei-maas" /home/oh-my-opencode-slim-huawei-maas
rm -rf "$TEMP_DIR"
| Scenario | Detection | Action |
|---|---|---|
| LiteLLM running | Health endpoint responds | Resolve master key, skip deploy |
| LiteLLM deployed but offline | Docker container exists (stopped) OR compose + .env exist at /home/LiteLLM-Huawei-MaaS-Proxy | docker compose up -d, resolve master key |
| No LiteLLM | No container, no files at canonical path | Extract from monorepo, init_env.sh, docker compose up -d |
In all scenarios, LITELLM_MASTER_KEY is required to mint the opencode virtual key. If not found in env/files, the user is prompted.
| # | Condition | Check | If Missing |
|---|---|---|---|
| 1 | bun | bun --version | https://bun.sh |
| 2 | jq | jq --version | https://stedolan.github.io/jq/ |
| 3 | Docker + Compose V2 | docker compose version | https://docs.docker.com/engine/install/ |
| 4 | git | git --version | install via package manager |
| 5 | python3 | python3 --version | install Python 3 |
| 6 | HUAWEI_MAAS_API_KEY | [ -n "$HUAWEI_MAAS_API_KEY" ] | export from Huawei Cloud console |
| 7 | HUAWEI_MAAS_EXTRA_API_KEYS (optional) | — | comma-separated extra MaaS keys for load balancing |
LITELLM_DIR=/home/LiteLLM-Huawei-MaaS-Proxy (canonical path)curl -sf http://127.0.0.1:4000/health/liveliness succeedsdocker compose up -d, wait for healthy, resolve master key$LITELLM_DIR (containers removed via compose down): docker compose up -d, wait for healthy, resolve master key$LITELLM_DIR, run scripts/init_env.sh --ci, docker compose up -d:4000, LITELLM_MASTER_KEY availabledocker compose logs litellm| Priority | Source | Location |
|---|---|---|
| 1 | Environment | $LITELLM_MASTER_KEY |
| 2 | .master-key file | $LITELLM_DIR/.master-key |
| 3 | .env file | $LITELLM_DIR/.env |
| 4 | Interactive prompt | User enters key |
command -v opencode succeedsbun install -g opencode~/.config/opencode/oh-my-opencode-slim.json existsbunx oh-my-opencode-slim@1.1.1 installsk- that passes a test completionscripts/mint-virtual-key.sh --no-budgetpreset == "LiteLLM-Huawei-MaaS" and council.presets definedassets/config/oh-my-opencode-slim.json.example to ~/.config/opencode/scripts/validate.sh — all checks must pass (0 failures)opencode — verify "LiteLLM-Huawei-MaaS" preset active in status bar<placeholder> in examples@ai-sdk/openai-compatible — NOT built-in openaiopenai/<model> format in LiteLLM provider (matching litellm_config.yaml)LiteLLM/openai/<model> (3-part)http://0.0.0.0:4000 (no /v1 suffix — SDK adds it; scripts use 127.0.0.1:4000 for curl)explore and general agents via agent (singular) with disable: true"lsp": true)max_budget)jq --arg for JSON substitution — never sedReplace placeholders in assets/config/opencode.jsonc.example using jq --arg:
jq --arg vk "$VIRTUAL_KEY" --arg mk "$MAAS_KEY" \
'.provider.LiteLLM.options.apiKey = $vk |
.provider["Huawei-MaaS"].options.apiKey = $mk' \
assets/config/opencode.jsonc.example > ~/.config/opencode/opencode.jsonc
| Placeholder | Source |
|---|---|
<LITELLM_VIRTUAL_KEY> | Virtual key from Step 5 |
<HUAWEI_MAAS_API_KEY> | $HUAWEI_MAAS_API_KEY env var |
Four presets in oh-my-opencode-slim.json — two via LiteLLM proxy (default), two direct to MaaS (debugging/fallback):
| Preset | Models | Route |
|---|---|---|
| LiteLLM-Huawei-MaaS (default) | All 5 | LiteLLM proxy → MaaS |
| LiteLLM-Huawei-MaaS-Lite | 3 (no v4-pro/v4-flash) | LiteLLM proxy → MaaS |
| Huawei-MaaS | All 5 | Direct to MaaS (no proxy) |
| Huawei-MaaS-Lite | 3 (no v4-pro/v4-flash) | Direct to MaaS (no proxy) |
Switch at runtime: /preset LiteLLM-Huawei-MaaS-Lite
| Agent | Model | Variant | Why |
|---|---|---|---|
| orchestrator | LiteLLM/openai/glm-5.1 | high | Strongest reasoning, delegation |
| oracle | LiteLLM/openai/deepseek-v4-pro | max | Deep thinking, architecture |
| council | LiteLLM/openai/deepseek-v4-pro | high | Independent deep review |
| librarian | LiteLLM/openai/deepseek-v3.2 | low | Highest RPM, cheapest, docs |
| explorer | LiteLLM/openai/deepseek-v4-flash | low | Fast, 1M context |
| designer | LiteLLM/openai/glm-5 | medium | General model, UI/UX |
| fixer | LiteLLM/openai/deepseek-v4-flash | high | Fast execution |
| Agent | Primary | Fallback |
|---|---|---|
| oracle | deepseek-v4-pro | glm-5.1 |
| council | deepseek-v4-pro | glm-5.1 |
| explorer | deepseek-v4-flash | deepseek-v3.2 |
| fixer | deepseek-v4-flash | glm-5 |
| Councillor | Model | Variant |
|---|---|---|
| alpha | LiteLLM/openai/deepseek-v4-pro | high |
| beta | LiteLLM/openai/glm-5.1 | high |
| gamma | LiteLLM/openai/deepseek-v3.2 | high |
| Service | URL | Credentials |
|---|---|---|
| LiteLLM Proxy | http://127.0.0.1:4000 | Virtual key (in opencode.jsonc) |
| LiteLLM Admin UI | http://127.0.0.1:4000/ui | Master key (from .master-key) |
| Prometheus | http://127.0.0.1:9090 | None |
| Grafana | http://127.0.0.1:3000 | admin / (from .env) |
| Config File | Location | Permissions |
|---|---|---|
| opencode.jsonc | ~/.config/opencode/opencode.jsonc | 600 |
| oh-my-opencode-slim.json | ~/.config/opencode/oh-my-opencode-slim.json | 600 |
| Symptom | Fix |
|---|---|
| LiteLLM won't start | docker compose ps, check port conflicts (4000, 5432, 9090, 3000), check .env |
| LiteLLM deployed but offline | docker compose -f /home/LiteLLM-Huawei-MaaS-Proxy/docker-compose.yml up -d |
| opencode won't start | jq . ~/.config/opencode/opencode.jsonc, check @ai-sdk/openai-compatible installed |
| Models not found | curl http://127.0.0.1:4000/v1/models, compare with litellm_config.yaml |
| Plugin not loaded | Add "oh-my-opencode-slim" to plugin array, re-run bunx oh-my-opencode-slim@1.1.1 install |
| 401 errors | Mint new virtual key, update opencode.jsonc |
| Wrong preset | Set "preset": "LiteLLM-Huawei-MaaS" or run /preset LiteLLM-Huawei-MaaS |
| "No presets configured" | Ensure council.presets is defined in oh-my-opencode-slim.json |
| Fallback not triggering | Set fallback.enabled: true, add chains for v4-model agents |
| Repo clone fails | Verify https://github.com/binrogithub/1-3-Cloud-Adoption-Skills is reachable |
| Requests timing out intermittently | Check request_timeout in litellm_config.yaml — should be 600s, not 10s; regenerate with ./scripts/generate_config.sh |
<HUAWEI_MAAS_API_KEY>, <LITELLM_VIRTUAL_KEY> placeholders in examples<prefix>...<suffix> in output — e.g. sk-abc123...xyz789jq --arg for substitution in scripts, never sed| Symptom | Cause | Fix |
|---|---|---|
| LiteLLM keeps restarting | DB not ready or wrong password | Check .env, docker compose logs db |
| 401 from proxy | Wrong master key or virtual key | Re-check key, mint new if needed |
| 404 model not found | Wrong model ID | Copy exact name from MaaS console |
| opencode won't start | Wrong provider npm package | Use @ai-sdk/openai-compatible, not openai |
| Budgets not decrementing | Model pricing is zero | Set non-zero token pricing in litellm_config.yaml |
| Virtual key expired | Duration or budget exceeded | Mint new key with --no-budget |
| Intermittent timeout errors from proxy | LiteLLM request_timeout too low (default was 10s, should be 600s) | Increase in litellm_config.yaml, regenerate with generate_config.sh, restart litellm |
http://127.0.0.1:4000provider key singular, agent key singularhttp://0.0.0.0:4000, valid virtual key, 5 models with openai/ prefixopenai/ prefix| Skill | Relationship |
|---|---|
| LiteLLM Huawei MaaS Proxy | Invoked — LiteLLM proxy deployment |
.
├── SKILL.md This file (agent-facing)
├── README.md Human-facing skill card
├── .gitignore Prevent secret commits
├── assets/
│ └── config/
│ ├── opencode.jsonc.example Template — copy to ~/.config/opencode/
│ └── oh-my-opencode-slim.json.example Template — copy to ~/.config/opencode/
└── scripts/
├── bootstrap.sh End-to-end orchestrator (idempotent)
├── install.sh opencode + plugin + config installer
├── mint-virtual-key.sh Mint scoped key from LiteLLM
└── validate.sh Validation checks
Use when analyzing or migrating Databricks CDC/Delta workflow scripts into a Huawei Cloud Chile MRS + OBS + Apache Hudi demo, including synthetic CDC data generation, notebook-triggered automation, Delta-to-Hudi replacement, deployment troubleshooting, and smoke validation. Also use when continuing the dockone ExampleApp demo package under the Codex outputs folder.
Configure Claude Code to use Huawei Cloud MaaS or ModelArts MaaS through an OpenAI-compatible endpoint, and optionally add Z.ai web-search-prime MCP search. Use when Codex needs to add a side-by-side claude-glm command that routes to Huawei MaaS glm-5.1 while preserving the original claude command on Anthropic, migrate claude itself to Huawei MaaS, install or configure claude-code-router, set API_KEY-based authentication, adjust context length, verify that Claude Code is actually backed by MaaS, or configure Z.ai MCP search with Z_API_KEY.
Deploy, configure, validate, troubleshoot, or extend an OpenAI-compatible API proxy backed by PostgreSQL, Prometheus, and Grafana, routing through Huawei ModelArts MaaS (ap-southeast-1) with multi-key load balancing. TRIGGER when the task involves LiteLLM proxy deployment, Docker Compose stack with litellm_config.yaml, Huawei MaaS model routing, virtual key or budget management, Prometheus/Grafana observability for LLM traffic, custom_callbacks.py TTFT/TPOT/ITL metrics, multi-key load balancing, or any reference to `LITELLM_MASTER_KEY`, `HUAWEI_MAAS_API_KEY`, or `docker compose` with this stack.
End-to-end screenshot-to-dashboard pipeline powered by MaaS vision models. Analyze a dashboard screenshot via OpenRouter free vision models, extract design tokens with anydesign, generate design.md + design-tokens.json, plan via CCPM, implement React+Ant Design+ECharts dashboard code, and iterate with screenshot comparison until ≥85% similarity. TRIGGER when the user provides a dashboard screenshot and wants to reconstruct it as a working web app, or says 'rebuild this dashboard', 'convert this screenshot to code', 'reproduce this BI dashboard', 'screenshot to React dashboard'. Do NOT use for: non-dashboard screenshots, Figma-to-code without screenshot, general UI development without a reference image.
Disciplined diagnosis loop for hard bugs and performance regressions. Reproduce → minimise → hypothesise → instrument → fix → regression-test. Use when user says "diagnose this" / "debug this", reports a bug, says something is broken/throwing/failing, or describes a performance regression.
Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".